forecasting operator: bug fix + automlx user specified metric based t…

…uning (#300)
oracle · Sep 11, 2023 · cef98e8 · cef98e8
2 parents 585abad + 9f829a3
commit cef98e8
Show file tree

Hide file tree

Showing 4 changed files with 23 additions and 5 deletions.
diff --git a/ads/opctl/operator/lowcode/forecast/const.py b/ads/opctl/operator/lowcode/forecast/const.py
@@ -25,4 +25,14 @@ class SupportedMetrics(str, metaclass=ExtendedEnumMeta):
     SMAPE = "smape"
 
 
-MAX_COLUMNS_AUTOMLX = 15
+AUTOMLX_METRIC_MAP = {
+    "smape": "neg_sym_mean_abs_percent_error",
+    "mape": "neg_sym_mean_abs_percent_error",
+    "mase": "neg_mean_abs_scaled_error",
+    "mae": "neg_mean_absolute_error",
+    "mse": "neg_mean_squared_error",
+    "rmse": "neg_root_mean_squared_error",
+}
+
+MAX_COLUMNS_AUTOMLX = 15
+DEFAULT_TRIALS = 10
diff --git a/ads/opctl/operator/lowcode/forecast/model/automlx.py b/ads/opctl/operator/lowcode/forecast/model/automlx.py
@@ -9,6 +9,7 @@
 import pandas as pd
 import numpy as np
 from ads.common.decorator.runtime_dependency import runtime_dependency
+from ads.opctl.operator.lowcode.forecast.const import AUTOMLX_METRIC_MAP
 from sktime.forecasting.model_selection import temporal_train_test_split
 from ads.opctl import logger
 
@@ -29,7 +30,8 @@ class AutoMLXOperatorModel(ForecastOperatorBaseModel):
         ),
     )
     def _build_model(self) -> pd.DataFrame:
-        import automl
+        from  automl import init
+        init(engine="local", check_deprecation_warnings=False)
 
         full_data_dict = self.full_data_dict
 
@@ -59,7 +61,10 @@ def _build_model(self) -> pd.DataFrame:
                 "" if y_train.index.is_monotonic else "NOT",
                 "monotonic.",
             )
-            model = automl.Pipeline(task="forecasting", n_algos_tuned=n_algos_tuned)
+            model = automl.Pipeline(task="forecasting",
+                                    n_algos_tuned=n_algos_tuned,
+                                    score_metric=AUTOMLX_METRIC_MAP.get(self.spec.metric,
+                                                                        "neg_sym_mean_abs_percent_error"))
             model.fit(X=y_train.drop(target, axis=1), y=pd.DataFrame(y_train[target]))
             logger.info("Selected model: {}".format(model.selected_model_))
             logger.info(

diff --git a/ads/opctl/operator/lowcode/forecast/model/neuralprophet.py b/ads/opctl/operator/lowcode/forecast/model/neuralprophet.py
@@ -9,6 +9,7 @@
 import numpy as np
 import optuna
 import pandas as pd
+from neuralprophet import NeuralProphet
 from torch import Tensor
 from torchmetrics.regression import (
     MeanAbsoluteError,
@@ -24,6 +25,7 @@
 )
 from ads.opctl import logger
 
+from ...forecast.const import DEFAULT_TRIALS
 from .. import utils
 from .base_model import ForecastOperatorBaseModel
 
@@ -166,7 +168,7 @@ def objective(trial):
                 )
                 study.optimize(
                     objective,
-                    n_trials=self.spec.tuning.n_trials if self.spec.tunning else 10,
+                    n_trials=self.spec.tuning.n_trials if self.spec.tuning else DEFAULT_TRIALS,
                     n_jobs=-1,
                 )
 

diff --git a/ads/opctl/operator/lowcode/forecast/model/prophet.py b/ads/opctl/operator/lowcode/forecast/model/prophet.py
@@ -15,6 +15,7 @@
 
 from ads.opctl import logger
 
+from ...forecast.const import DEFAULT_TRIALS
 from .. import utils
 from .base_model import ForecastOperatorBaseModel
 
@@ -144,7 +145,7 @@ def objective(trial):
                 )
                 study.optimize(
                     objective,
-                    n_trials=self.spec.tuning.n_trials if self.spec.tunning else 10,
+                    n_trials=self.spec.tuning.n_trials if self.spec.tuning else DEFAULT_TRIALS,
                     n_jobs=-1,
                 )